libstdc++/9378: 64-bit g++: construct base:std::moneypunct_byname<char>->memory fault

[Martin Sebor]

Benjamin Kosnik wrote:
> On Wed, 29 Jan 2003 14:24:10 -0700
> Martin Sebor <sebor@roguewave.com> wrote:
> 
> 
>>FWIW, I don't think the requirements on the ctors of class
>>locale are relevant in this case. The problem is, IMO, that
>>there are no requirements at all on the ctors of the byname
>>facets, so pretty much anything goes. I.e., abort is just
>>as legal as successful construction.
> 
> 
> Hmm. I wasn't paying attention, and didn't realize these were _byname facets. 
> It doesn't really matter though, because....
> 
> 22.1.1.1.2 - Class locale::facet [lib.locale.facet]
> 
> -4- For some standard facets a standard ``..._byname'' class, derived
> from it, implements the virtual function semantics equivalent to that
> facet of the locale constructed by locale(const char*) with the same
> name. 
> 
> So, with this equivalence argument, I think the original point still
> holds: these name strings are implementation-defined, and people should
> not assume that any arbitray string will in fact, lead to a constructed
> locale object.

I agree (with the name argument). I would go even further and
assert that programs that directly construct any of the _byname
facets have unspecified behavior. The only safe (or portable)
way to use the _byname facets is via a call to use_facet<
xxx_byname>(), and even that only after a corresponding call
to has_facet<xxx_byname>()).

In fact, I would even propose that we remove the _byname ctors,
or better yet, the entire facets from the standard ;-) (only
half joking).

Regards
Martin